● Rottawhite — AI Systems Studio

Loss runs, normalised across every carrier format.

Every carrier formats loss runs differently, none of them formats them well, and an underwriter needs five years of them in one view.

Test it in a week — $2,500 →How we prove accuracy →

Today

What the desk looks like now

  • Loss runs are requested by email and chased for weeks, delaying the submission they were needed for.
  • They arrive as PDFs, sometimes as scans of printouts, in a different layout from every carrier.
  • Someone retypes them into a spreadsheet to get a five-year view, and the retyping is where the errors enter.
  • Analysis that should inform pricing happens late, or does not happen.

The system

What the agent does

Requests and tracks

Generates carrier requests, tracks what is outstanding, and chases on a schedule so the follow-up does not depend on someone remembering.

Parses any format

Extracts claim-level detail — dates, cause, paid, reserved, status — from whatever layout the carrier uses, including scanned output.

Normalises to one schema

Different carriers naming the same field differently is a mapping problem, and mapping is exactly the kind of thing worth doing once properly.

Produces the view underwriting wants

Loss ratio by year, frequency and severity trends, large-loss detail, and open reserves — assembled rather than retyped.

Evidence

What gets logged

The audit trail is the part that makes this usable in regulated work. Every output traces back to a document.

  • Every claim record traced to its source document and page
  • Every field mapping applied, so a normalisation choice can be inspected
  • Original values retained alongside normalised ones
  • Request and chase history per carrier
How the audit trail works →

How accuracy is measured

Built from your own documents, including the bad ones. An average that hides the hard cases is not a number worth having.

  • Claim-level extraction accuracy against labelled loss runs from your carriers
  • Totals reconciliation — extracted figures must sum to the document totals
  • Accuracy on scanned and low-quality documents specifically, since that is where extraction fails
How the eval harness works →

Questions

Common objections

Some of our loss runs are scans of faxes. Really?

Really, and that case belongs in the test set. Accuracy on poor scans is lower than on clean PDFs, which is why it is measured separately rather than hidden inside an average.

How do you know the extraction is right?

Beyond the labelled test set, extracted claim figures are reconciled against the totals printed on the document. A mismatch is an automatic exception.

Also

Other back-office workflows

Workflow

Submission intake

A submission arrives as a forwarded email with six attachments in four formats. Someone keys it in. That someone is usually expensive.

Workflow

Policy checking

The carrier sends the policy. Somebody is supposed to read all sixty pages and confirm it matches what was agreed. Often nobody does.

Workflow

Certificates of insurance

Low value per unit, enormous volume, zero tolerance for error, and it never stops. Certificates are the definition of work a machine should do.

Next step

Start with one workflow.

A week, a fixed fee, and a measured answer on your own documents. If it will not work, you find out for $2,500.

Book a 30-min call →The $2,500 sprint →